> the majority opinion seems to be that training diffusion models on copyrighted work is totally fine

Well, maybe they were all just downvoted into invisibility. But up to your post, I have seen none.